Re: 32bit kernel is broken for Linux-5.1-rc1 due to GCC cmodel=medlow

From: Anup Patel <anup@brainfault.org>
To: Christoph Hellwig <hch@infradead.org>
Cc: Damien Le Moal <Damien.LeMoal@wdc.com>,
	Palmer Dabbelt <palmer@sifive.com>,
	Anup Patel <anup.patel@wdc.com>,
	Mike Rapoport <rppt@linux.ibm.com>,
	Atish Patra <atish.patra@wdc.com>,
	linux-riscv@lists.infradead.org
Subject: Re: 32bit kernel is broken for Linux-5.1-rc1 due to GCC cmodel=medlow
Date: Sun, 24 Mar 2019 07:12:07 +0530	[thread overview]
Message-ID: <CAAhSdy07GsHXpG9MJZ4aiB5uG7D+9RA+17Nb1MrWdFg9ee6rUA@mail.gmail.com> (raw)
In-Reply-To: <20190323172153.GA9085@infradead.org>

On Sat, Mar 23, 2019 at 10:51 PM Christoph Hellwig <hch@infradead.org> wrote:
>
> On Fri, Mar 22, 2019 at 07:15:37PM +0530, Anup Patel wrote:
> > >
> > > I am not saying the revert is the final solution.  I'm just saying
> > > that 5.1-rc is not the place we should be doing the major surgery.
> > > Revert fox 5.1, fix for real for 5.2+.
> >
> > If we revert for time being then even fixmap support has to be
> > reverted.
> >
> > The patch only updates access to kernel symbols via __load_pa()
> > and __load_va(). There is no functional change as such. The patch
> > is tested on QEMU/virt 64bit, QEMU/virt 32bit, and SiFive Unleashed.
>
> Just curios - if you just move setup_vm and dependencies into a new
> file that is compiled with -mcmodel=medany, just like setup.c,
> what issues would we still have that require explicitly passing the
> load address?

There are four cases here:
1. 64bit with cmodel=medany (default for 64bit system)
2. 64bit with cmodel=medlow
3. 32bit with cmodel=medany
4. 32bit with cmodel=medlow (default for 32bit system)

I had tried using cmodel=medany as default for both 64bit and 32bit system
and it works fine because there is no issue with cmodel=medany but it is a very
crude fix and things break again if someone explicitly selects cmodel=medlow
using kconfig.

The patch for using __load_va()/__load_pa() accessor in setup_vm()
solves the problem for all four cases above.

If you still think the patch is too big as rcX fix then alternate approach is:
1. For Linux-5.1-rcX, we use cmodel=medany by default for both 32bit
and 64bit RISC-V (I can send a patch for this immediately)
2. Consider __load_va()/__load_pa() patch for Linux-5.2

Regards,
Anup

_______________________________________________
linux-riscv mailing list
linux-riscv@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-riscv